How reliably can we estimate trends of surface weather extremes? A conceptual study using ERA5 reanalyses

How the frequency and intensity of extreme weather events is affected by global warming in different regions is one of the central questions of climate change research, with obvious direct implications for climate change adaptation. A standard approach of defining weather extremes is to consider the exceedance of a percentile threshold, calculated from the statistical distribution of a meteorological variable of interest in a predefined reference period. Trends can then be assessed by considering the frequency of threshold exceedances in a period that extends beyond the reference period. While this approach appears rather straightforward, it comes with several choices related to the parameter, percentile threshold, aggregation period, reference period, and boosting interval. Here aggregation period refers to the question whether, e.g., precipitation extremes are considered with a duration of 1 hour or 1 day or multiple days, and the boosting interval is the symmetric time window used to calculate percentiles for a given day of year. When checking these partly methodological choices in previous studies, e.g., those referenced in the IPCC report, it becomes evident that different studies made different choices. Since there is no obvious “best choice”, it is important to quantify the influence of these choices on the resulting trend estimates. Therefore, this study uses ERA5 reanalysis data to systematically and globally explore the trends in 2-m temperature (T2m) and precipitation (P) and their robustness with respect to the aforementioned parameters. Key results are that (i) trends vary strongly between regions, (ii) they are methodologically more robust for T2m than for P, (iii) in regions with weak P trends, the sign of the trend depends on the methodological choices. These explorative analyses with ERA5 data are complemented by synthetic data experiments, in particular to investigate the influence of the boosting window. We suggest that trend analyses of percentile threshold exceedances of any parameter in any dataset should consider these methodological sensitivities in order to communicate robust estimates.